Dynamic Bilateral Trading in Networks

Transcription

1 Dynamic Bilateral Trading in Networks Daniele Condorelli November 2009 Abstract I study a dynamic market-model where a set of agents, located in a network that dictates who can trade with whom, engage in bilateral trading for a single object under asymmetric information about the private values. My equilibrium characterization provides new insights into how economic networks shape trading outcomes. Traders who link otherwise disconnected areas of the trading network become intermediaries. They pay the object at their resale values but, if they have a high value, they consume and extract a positive rent. All other traders, except for the initial owner of the object, make zero profit. The object travels along a chain of intermediaries before someone consumes it. Intermediaries who are located later in the trading chain have a lower probability of acquiring the object, but they pay lower prices for it. Compounding, early intermediaries gain a payoff advantage over late ones. Adding links to the network increases downstream competition and it is beneficial to the initial owner. However, it has ambiguous effects on the other traders and may be detrimental to total welfare, when information is asymmetric. More generally, inefficient outcomes are possible if information is not complete and the network is not fully connected. Special thanks go to Andrea Galeotti and Philippe Jehiel who read several earlier drafts and provided suggestions that improved the paper substantially. I also thank Yeon-Koo Che, Rakesh Vohra and Asher Wolinsky for several insightful conversations. Finally, I thank all seminar participants at Northwestern University for their helpful comments before, during and after my presentation. Department of Economics at University College London (Ph.D. candidate) and Department of Managerial Economics and Decision Sciences at Northwestern University (visiting scholar). Contact address: 2001 Sheridan Road, Evanston, IL 60208, USA. Telephone: northwestern.edu/faculty/condorelli/ 1 Website:

2 1 Introduction I consider a finite-horizon dynamic model where a set of agents located in an exogenous network engage in bilateral trading for a single object, initially owned by one of them. Traders are risk neutral and have a high or low monetary evaluation for the object, which is private information. Values are drawn from the same support and are independently but possibly asymmetrically distributed. In each round the owner of the good can either consume the good, make a take-it-or-leave-it offer to a single agent located in his neighborhood, or wait one round. If an offer is made and accepted, the terms of trade are enforced. Thereafter, a new round of trading starts. The game ends if someone consumes the object or a known deadline is reached. Traders do not discount the future and all actions are publicly observed. This model provides a novel framework to investigate how the existence of a network of potential transactions, in which agents are embedded, affect their trading and bargaining outcomes under asymmetric information. In particular, the model captures well the main features of over-the-counter (OTC) trading in financial instruments. OTC refers to bilateral trading between buyers and sellers, as opposed to trading in centralized exchanges. 1 Products such as credit default swaps, forward rate agreements, and exotic options are almost always traded in this manner. These instruments are often associated with a specific maturity date and can change hands several times before reaching a consumer. Furthermore, bilateral transactions are not guaranteed by the stock exchange, and are subject to counter-party risk. Therefore, bonds of trusts between trading firms are particularly important, and represent the main source of the trading network structure in which OTC transactions take place. 2 A number of insights emerge from the characterization of a set of perfect Bayesian equilibria for the market game outlined above. Two types of active traders arise endogenously in each equilibrium, final customers and intermediaries. 3 Final customers receive only offers at a high price, that they accept only when they have a high value for the object, and leave them with zero profit. Intermediaries acquire the object at lower prices, equal to their resale values. They resell the object in their neighborhood if their value is low, while they consume 1 The notational amount of OTC derivatives outstanding at the end of 2008 exceeded 500 trillion dollars, according to statistics from the Bank for International Settlements. 2 See Allen and Babus (2009) for a survey of papers on financial networks. This model could be applied also to international trading networks (e.g. see Rauch (1999), (2001) and Casella and Rauch (2002)) 3 A trader is active if he takes at least one action with positive probability along the equilibrium path. 2

3 it and make a positive profit, if their value is high. Whether a trader becomes an intermediary, a final customer or remains inactive is determined jointly by the prior information and the network structure. Traders on the periphery of the network and high value traders tend to become final customers, while players that are essential to provide access to valuable areas of the network or have a low expected value become intermediaries (see subsection 5.1). This is in line with a large body of empirical evidence illustrating that bridging two areas of an economic network that would otherwise be disconnected provides a payoff advantage. See for example the analysis of structural holes in Burt (1992). In equilibrium, the object travels along a chain of intermediaries who make offers to final customers, until consumption takes place. Intermediaries acquiring the object later have a smaller probability of realizing a profit upon consumption. Though, the price that they pay is lower, because the resale values decrease as rounds pass, every time an offer is rejected. In general, the first effect dominates the second, and intermediaries who acquire the object earlier in the trading chain are better off than those who acquire it later (see subsection 5.2). 4 Ex-post efficiency is attained under complete information (see section 3), when the network is fully connected, or if players are ex-ante identical. However, in general, the interplay between asymmetric information and network structure can generate inefficiencies (see subsection 5.4). In fact some high value player might remain inactive and never receive an offer, even if all other traders have a low value for the object. Comparative statics show that the initial owner is better off when the network becomes more connected. More generally, an increase in connectivity will increase, ceteris paribus, the resale value of each trader. However, for traders other than the initial owner, an increase in the resale value has ambiguous effects. It may harm an intermediary, because it increases the price that he has to pay for the object. It may be beneficial to a final customer, if it changes him in to an intermediary. Moreover, I show with an example that an increase in connectivity can also represent a source of inefficiency (see subsection 5.3). 5 This paper is a first attempt to introduce asymmetric information into a dynamic model 4 Intermediaries can also receive other offers earlier in the game, before they acquire the good for resale, at a price that they will only accept if they have a high value. When this is the case, a clear-cut payoff ranking, favoring early over late intermediaries, is available only if traders are ex-ante identical. 5 In addition, subsection 6.1 analyzes networks with a large number of traders and subsection 6.2 characterize optimal networks. 3

4 of exchange in networks. Its main contribution is to the literature on the impact of network structures on market outcomes, a topic which has not been explored exhaustively. 6 Kranton and Minehart (2001) was perhaps the first paper to provide a non-cooperative model of exchange in a network. They study networks of buyers and sellers under asymmetric information. Sellers use auctions to sell their goods, after which no further trade takes place. Therefore, intermediation and resale do not figure in their model. Blume et al. (2009) study a two stage game of complete information, where intermediaries, sellers and buyers coexist. Intermediaries choose bid and ask prices to offer to sellers and buyers to whom they are connected. Traders accept or reject their offers. Nava (2009) develops a static model of Cournot competition in networks, under complete information. Players who buy sell and retail are determined endogenously in equilibrium. In his model prices increase along the supply chain because each trader has local market power. 7 In contrast with these three papers, the approach to trade in my paper is dynamic and the market model is fully decentralized (i.e. only bilateral negotiations are allowed). The most closely related dynamic model is Gale and Kariv (2007). They analyze trade in a network under complete information. Their main result, that an efficient outcome is attained in the long run, is in line with my full information finding in proposition 1 (see section 3). 8 The present paper is also related to the literature on decentralized markets with matching and bargaining under complete information, initiated by Rubinstein and Wolinsky (1985). Calvo-Armengol (2003) and Corominas-Bosh (2004) were the first to consider an environment where the matching technology is constrained by a network structure. 9 In contrast to this set of papers, in my paper the analysis is not restricted to bipartite buyer-seller networks, and there is both incomplete information and resale of goods in equilibrium. 6 As Jackson (2008) puts it in a section on networked markets [...] there is much left to be learned in this extensive area of application where network play such a central and critical role. 7 Kakdade et al. (2005) and (2004) adopt a static and centralized competitive equilibrium perspective. In their models traders are price takers and prices are defined by market clearing conditions. 8 My work is also related to Jehiel and Moldovanu (1999). They study dynamic resale processes with externalities under complete information, in a setting where everyone can trade with everyone else. Zheng (2002) and Calzolari and Pavan (2006), on the other hand, analyze models including resale under asymmetric information. Their focus is non mechanism design and their informational structure is more general. However, they do not address the limitations imposed by a network structure. 9 Recent contributions to this area include Polanski (2007a), Abreu and Manea (2009) and Manea (2009). Furthermore, Polansky (2007b) considers a model of information pricing in networks, with random matching. 4

5 2 Model The economy consists of a set of traders, N {1,..., n}, and two types of goods: money, which is distributed to all players in large quantity, and a single indivisible object, initially owned by agent 1. Each trader i is risk neutral and has a binary private monetary evaluation for the good, v i, normalized to be either zero or one. Therefore, if trader i consumes the good with probability x i and expects to pay m i, then his utility is x i v i m i. The common prior probability that v i = 1 is denoted π i [0, 1], and values are assumed to be independently distributed. Traders are located in a connected and undirected network G = (N, E), where the set of players coincides with the set of vertices N, and E 2 N N is the set of edges between pairs of players. Existence of an edge between two players means that trade between them is possible. Traders i and j are neighbors if {i, j} E (also written ij E). A network is undirected if ij E implies that ji E. A path from i to j in G is a non-empty graph where the set of vertices is {i, b 1,..., b m, j} N and the set of edges is {ib 1, b 1 b 2,..., b m j} E. I refer to a path by the ordered sequence of its vertices b(i, j) = (i, b 1,..., b m, j), and refer to the length of the path by the cardinality of the set of edges. A network G is connected if there is a path between every pair of players. 10 The game consists of a finite number of rounds T. Agents do not discount the future and all actions taken by all traders are observed by everyone in the network. Each round t develops in a number of stages. Denote by s t N the owner of the good at the beginning of round t. First, s t can make a take-it-or-leave-it offer to one of his neighbors, or make no offer. In the former case, I denote the chosen neighbor i t and the price asked p t [0, 1]. Second, if s t makes offer (i t, p t ), then i t decides whether to accept or reject it. If the offer is accepted, s t transfers the object to i t and receives a payment p t from i t. Finally, at the end of the round, the current owner of the object decides whether or not to consume it. If the object is consumed, the game ends. Otherwise, the game proceeds to round t + 1. The game also ends if it reaches the end of round T. Of course, mixed strategies are allowed at all stages. Everything but private values is common knowledge, including the network structure. 10 The assumption of connectedness is without loss of generality. No activity occurs in an area of the network that is disconnected from the initial owner of the object. 5

6 A triple G, π, T is a network trading game, which is an extensive form games with observed actions, independent types and a common prior. 11 The adopted solution concept is perfect Bayesian equilibrium. Informally, a perfect Bayesian equilibrium is a strategy profile and a belief system such that the strategies are sequentially rational given the belief system, and the belief system is consistent with Bayesian updating, wherever possible, given the strategy profile. See Fudenberg and Tirole (1991) for a formal definition. 12 Some of the assumptions I make can be relaxed. First, most of the results extend to the case in which, in case of sale, the seller must bear a transaction costs, which is edge specific and depends on the identity of the buyer and the seller. From here on, therefore, it is intended that each result, unless otherwise specified in a footnote, will be valid in a setting with transaction costs, perhaps after minor modifications. Second, discounting can be accommodated easily within a finite horizon model, and, again, most of the result would hold with only slight modifications. Third, full observability of prior actions is not strictly necessary for the survival of the specific equilibria that I construct. In fact, because generically each player play a pure strategy along the equilibrium path, knowing the round of the game in which an offer is received and the identity of the seller making that offer will provide a sufficient statistic of all past actions. 3 Efficiency and Complete Information Games This section develops the benchmark case of trading games played under complete information. As a preliminary step, I present the standard notion of efficiency tailored to this environment. If the profile of values is known, a feasible outcome is an allocation of the goods 11 This is a Bayesian game of incomplete information and, for each player, the only two types à la Harsanyi are the payoff types, either zero or one. 12 The notion of Perfect Bayesian Equilibrium (PBE), which is extensively used in economics, imposes as assumptions about the belief systems a number of propositions which appear in Kreps and Wilson (1982) treatment of Sequential Equilibirum (SE) as consequences of a single consistency requirement. I adopt the notion of PBE for two reasons. First, there are some technical difficulties in extending the sequential equilibrium notion to infinite games. Second, the set of SE and the set of PBE would coincide in my environment with only two possible types if the set of possible prices was discrete (see Fudenberg and Tirole (1991)). Therefore I conjecture that, if SE was appropriately defined, the two sets would coincide also in the limit (see Fudenberg and Levine (1986)). 6

7 to the players that is achievable within T rounds of sequential trade. An outcome is Pareto efficient (or simply efficient) if it is feasible and there is no alternative feasible outcome that would make at least one trader strictly better off without making any other trader worse off. Because utility is linearly transferable through monetary exchanges, an efficient outcome is one where the object is allocated to a value one player, if any, who can be reached by the object in at most T rounds of trade. The first and second fundamental theorems of welfare economics hold trivially in this economy, since competitive equilibrium outcomes and efficient outcomes coincide. 13 The next proposition states that, even when trading is decentralized, an efficient outcome is always attained under complete information. The proof is in appendix B. Proposition 1 (Equilibrium in the Full Information Game). Assume that values are known, π i {0, 1} for all i N. In every network trading game G, π, T every sub-game perfect equilibrium outcome is efficient. The equilibrium takes the following form. Player 1, also hereinafter referred to as the initial owner, consumes the good if v 1 = 1, or no other player with has value one can be reached in less than T rounds of trade. Otherwise, denote by O T the set of traders other than 1 who have value one and can be reached in less than T rounds. For each i O T and, for each path (1, b 1,..., b m, i) with length less than T, there is a subgame perfect equilibrium that takes the following form. Trader 1 sells the good to b 1, b 1 sells the good to b 2 and so on, until player i O T buys the good from b m and consumes it. The price along the trading path is constant and equal to one. Every trader other than the initial owner makes zero profit The main insight from the analysis is that a network structure does not generate inefficiency per se, if information is complete. This is in line with other work on networked economies in settings without informational asymmetries (e.g. Blume et al. (2009) and Gale and Kariv (2007)) Competitive equilibria outcome can be computed in the standard way, allowing for the technical limitations imposed by the network structure. 14 In general, as Gomes and Jehiel (2005) analysis implies in a related environment, in the absence of consumption externalities, dynamic processes of social and economic interaction under complete information tend to converge in the long run to an efficient outcome. 7

8 4 Trading Equilibria Let us now focus on the case where information on values is asymmetric. The next theorem is the first main result of this paper. Theorem 1 (Equilibrium Existence). A perfect Bayesian equilibrium exists for each network trading game G, π, T. 15 The proof is constructive and follows the backward induction logic (see appendix A, where transaction costs are explicitly considered). It proceeds in three steps. First a unique equilibrium for a game starting in the last round T is constructed, for every possible owner in T and for any state of beliefs. Second one or more equilibria for an arbitrary game starting in round t 1 are constructed for each possible owner and profile of beliefs, assuming that the set of equilibria for a game starting in t has been computed. Finally, an equilibrium for the whole game is constructed by induction. Observe that equilibrium uniqueness can not be guaranteed generically, unless the set of possible networks is appropriately restricted. The next example clarifies the equilibrium construction algorithm. The exposition includes a number of observations related to the general properties of trading equilibria, which are instrumental to the discussion in section 5. Example 1 (Equilibrium Construction). Players {1, 2, 3, 4} are located in a network, where E = {12, 13, 23, 24} and π = {0, π, 1/8, 1/2} (see Figure 1). Assume that T = 3 and that player 1 is the initial owner. The game is solved backward, starting with round T. Round T. Take any arbitrary history that led to round T, and assume that µ T is the state of beliefs and the owner is player 1, that is s T = 1. Since T is the last round, both players 2 and 3 will accept every price p T 1 if they have value one, while they will reject every price greater than zero if they have value zero. Therefore, in equilibrium, player 1 offers the object at price p T = 1 to player 3 if µ T 3 > µ T 2. Otherwise, he offers the object at p T = 1 to player 2. Player 1 s expected payoff is max{µ T 3 ; µ T 2 } and all other players make zero profit. An analogous argument can be made to obtain an equilibrium when s T is either 2, 3 or Equilibrium existence,to my knowledge, is not guaranteed by existing results because the game is dynamic and the action space is not finite. 8

9 1 Π Π 2 Π 3 Π Π Figure 1: Trading Network in Example 1 Round T-1. Let µ T 1 be the state of beliefs at the beginning of round T 1 and assume first that s T 1 = 1. Hereafter, I can assume that µ s t = 0, because if v s t = 1, then s t would have consumed the good instead of putting it up for sale. In general, I can make the following two observation. Observation 1. In any round t, it is a dominant strategy for a trader i who acquires the good, to consume it only if v i = 1, and to put it up for sale only if v i = 0. Observation 2. It follows from Bayesian updating that, whenever a trader puts the good for sale, he must have value zero. That is, µ t s = 0 for all t and s t. t Player 1 can make an offer to either 2 or 3, or make no offer, in which case he remains the owner in round T and beliefs do not change. To obtain an equilibrium it is necessary to pin down the acceptance strategies of players 2 and 3, for any price that may be offered 9

10 to them. The analysis in round T shows that rejecting an offer in T 1 will provide zero utility to 2 and 3, because the owner in round T will again be player 1. Therefore, if they have value one, it is a best reply for 2 and 3 to accept any price p T 1 1. If they have value zero, a best reply for 2 and 3 is to accept any price below or equal to their value from reselling the object in period T. The resale value of a player i in round t, denoted V t i, depends on the network configuration and on the state of beliefs. If player i accepts an offer in round T 1 and puts the object up for sale, he is signalling a value zero and therefore µ T i = 0. The beliefs about the seller and the other players remain unchanged. Therefore V2 T = max{µ 3 T 1 ; µ 4 T 1 } and V3 T = µ T 2 1. This logic applies throughout the game, as stated in the two observations below. Observation 3. In equilibrium a trader i with v i = 0 accepts an offer in round t if and only if the price offered p t is equal to or below his resale value in round t + 1, denoted V t+1 i. Observation 4. In equilibrium, a buyer i with v i = 1 will always accept any price p t V t+1 i, or otherwise he will reveal his type and in equilibrium will make zero profit for sure. Having fixed the strategies of the potential buyers, the following simple observations apply to the strategies for sellers. Observation 5. If i s acceptance probability is constant in the interval [p l, p h ], then it is never optimal for a seller to offer the object to i at any price p t [p l, p h ] such that p t p h. Observation 6. In any round t, it can never be optimal for a seller s t to offer the object too a trader i at a price below i s resale value, that is V t+1 i. Therefore, if player 1 is the owner in round T 1, he only needs to consider four possible offers, in addition to not making any offer: (i) Player 1 asks p T 1 = 1 to player 2. In this case player 2 accepts if and only if v 2 = 1. In case of a refusal player 1 assumes that µ T 2 = 0 and in round T offers p T = 1 to player 3. The expected payoff for player 1 is: µ T (1 µ T 1 2 )µ T 1 3. The expected payoff for 2 and 3 is zero. (ii) Player 1 asks p T 1 = 1 to player 3. In this case player 3 accepts if and only if v 3 = 1. In case of a refusal player 1 offers p T = 1 to player 2 in round T. The expected payoff for player 1 is: µ T (1 µ T 1 3 )µ T 1 2. The expected payoff for 2 and 3 is zero. 10

11 (iii) Player 1 asks p T 1 = V2 T to player 2. Player 2 accepts for sure and his expected profit is 1 V2 T if he has value one and zero otherwise. Player 1 obtains V2 T. (iv) Player 1 asks p T 1 = V3 T to player 3. In this case player 1 obtains V3 T. Player 3 obtains 1 V3 T if v 3 = 1 and zero otherwise. Player 1 is indifferent between offers (i) and (ii). Moreover, offer (iv) is always dominated by offer (i) and (ii). Finally, not making any offer is also strictly dominated by (i) and (ii). Therefore, an optimal strategy for player 1 is either (iii) or either of (i) and (ii), depending on the state of beliefs. To sum up, an equilibrium is obtained for the continuation game starting in T 1 with s T 1 = 1. Analogously, an equilibrium can be obtained when the seller at T 1 is another player than 1. Round T 2 = 1. Let us now analyze the entire game. Therefore, let µ 1 = π and s 1 = 1. The analysis is developed under two assumptions. Under assumption A the expected value of player 2, π 2 = π, is relatively low, while under assumption B it is relatively high. Assumption A: 1/2 > 1/8 + 7/8π. First, we need to compute the equilibrium acceptance strategies for traders 2 and 3. Consider 2 first. If v 2 = 0, according to observation 3, player 2 accepts every price below or equal to V T 1 2 = 1/2 + 1/16 = 9/ If v 2 = 1, by accepting an offer at a price p 1 he gets 1 p 1, whereas if he rejects the offer he gets a payoff that can be computed by looking at round T 1, for the case where 1 is the owner in T 1, µ T 1 2 is determined according to Bayesian updating and µ T 1 2 = µ T 2 2. Let V2 T 1 (µ T 2 1 ) be this payoff. In this case V T 1 2 (µ T 1 2 ) = 1/2 for all µ T 1 2 π 2, because it is always optimal in round T 1 for player 1 to offer player 2 a price of 1/2. In fact, under assumption A, 1/2 > 1/8 + 7/8π, which is the maximum that player 1 could achieve by making a different offer in T 1. Therefore an optimal strategy for player 2 is to accept every p 1 9/16 and reject higher prices. Next, consider player 3. If v 3 = 0, player 3 accepts every price below or equal to his resale value V T 1 3 = max{π, 1/2} = 1/2, where the second equality follows from assumption A. 17 If 16 This is the expected payoff that player 2 obtains in the game starting at round T 1, with s T 1 = 2 and µ 2 2 = π 2, µ 2 2 = 0. In this subgame player 2 resells at price one to 3 and This is the expected payoff that he will obtain in the game starting at round T 1, with s T 1 = 3 and µ 2 3 = π 3, µ 2 3 = 0. In this subgame trader 3 asks price 1/2 to 2, who then resells to 4 at price one. 11

12 v 3 = 1, V T 1 (µ T 1 3 ) = 0 for any µ T 1 3 because player 3 will obtain no offer in round T 1. Therefore, it is optimal for player 3 with v 3 = 1 to accept every price p 1 1. Equilibrium path under assumption A. In round one player 1 offers p 1 = 1 to player 3, who accepts only if v 3 = 1, and thereafter consumes. In case of a refusal, player 1 offers p 2 = 1/2 to player 2, who accepts regardless of his value, and consumes if v 2 = 1. If v 2 = 0, in the last round, player 2 asks for a price p 3 = 1 to player 4, who accepts only if v 4 = 1. Assumption B: 1/2 < π. The acceptance strategy of player 3 is computed as under Assumption A. The only difference is that the resale value of 3 will now be equal to π. Instead, when v 2 = 1 player 2 will not play a pure strategy after an history in which he obtains an offer in round T 2. To see this, note that, under assumption B, the best strategy for player 1 in round T 1 if he believes that player 2 has value one with probability π, is to ask p T 1 = 1 from player 3 and, if refused, to ask p T = 1 from player 2; hence, V T 1 2 (π) = 0. On the other hand, if player 1 in round T 1 believes that player 2 has value zero, when in reality player 2 has value one, the expected payoff of 2 is V T 1 2 (0) = 1/2 > V T 1 2 (π). It follows that it can not be part of an equilibrium for player 2 with v 2 = 1 to mimic a value zero trader and refuse all prices above V T 1 2 = 9/16. If he adheres to this strategy he obtains zero profit, whereas accepting a price p (9/16, 1) would provide a positive payoff. Moreover, it cannot be part of an equilibrium for 2 with v 2 = 1 to accept with probability one offers above 9/16, because if he rejects, then the seller would believe that he has value zero and 2 would find this profitable. 18 Therefore, in this case the acceptance strategy of player 2 with v 2 = 1 must involve mixing. In particular, player 2 accepts any price below or equal to 9/16, and he randomizes his acceptance decision for all prices in (9/16, 1], in such a way that, upon refusal, the seller changes his belief about 2 to µ T 1 2 = 3/7. However, for this to be a best reply player 2 must be indifferent between accepting and rejecting an offer at a price p 1 in that interval. Therefore, the payoff of player 2, upon refusing price p 1, must be equal to 1 p 1. This can happen because if player 1 assumes µ T 1 2 = 3/7, then he is indifferent between two courses of action starting in T 1: either asking price p 2 = 1/2 from 2, or asking price p T 1 = 1 from 3 and in case of refusal asking price p T = 1 from 2 in round T. Therefore, he can randomize 18 Note that player 2 cannot refuse a price below or equal to 9/16 as otherwise he would signal that he has value one. 12

13 between the two options, as a function of the price posted to 2 in T 2 in such a way that player 2 becomes indifferent between accepting and rejecting at T Equilibrium path under assumption B. Player 1 will ask a price of one from player 3 in round T 2, who will accept only if he has value one. In case of a refusal player 1 will ask a price of one from player 2 in round T 1. In round T 1 player 2 will accept any price below or equal to 1. It should be emphasized that the outcome of the game under assumption A is ex-post efficient, while it is inefficient under assumption B. In fact, in the latter case, player 4 never gets an offer, even if he happens to be the only player in the game with value one. Therefore, in contrast to the complete information case, ex-post efficiency is not guaranteed under incomplete information. This point is discussed in more detail in section 5.4. Even though in the above example it is assumed that T = 3, increasing the number of rounds to four or more will not change the set of equilibrium payoffs. The idea is that, because learning is irreversible, the number of payoff relevant offers that can be made in any equilibrium of any game, before everyone learns everything, is finite. Therefore, the equilibria constructed according to Theorem 1 are not sensitive to the time horizon of the game, when this is sufficiently long. 20 Proposition 2 (Time Stability of Equilibria). For each given G and π there exist a number of rounds T such that, for each T T the set of equilibrium payoffs in the game G, π, T, computed according to the equilibrium construction algorithm in appendix A, coincides with the set of equilibrium payoffs in the game G, π, T. This property is exploited in the next section. I refer to it either by using T rather than T in the definition of a network trading game, or by stating that the number of rounds is sufficiently large. 19 This example also shows that Markov equilibria will not always exist in network games, as sometimes the seller s future price will have to depend on the present one, which does not affect the continuation payoffs in the game that starts in the following round. 20 This property is common to other settings with a fixed deadline and no discounting (see e.g. Jehiel and Moldovanu (1999)). 13

14 5 Equilibrium Analysis This section characterizes the main properties of equilibrium outcomes. From hereon, I focus on equilibria where sellers play pure strategies along the equilibrium path. Such equilibria always exist, while equilibria where sellers randomize their decisions along the equilibrium path are non-generic. In fact, they can be eliminated by introducing transaction costs and thereafter appropriately restricting the set of possible priors and transaction costs, without reducing the dimensionality of the set where these parameters are defined. 21 In what follows, I discuss first how the location of a trader in the trading network determines the role that he will play in the trading process and the terms of trade he will face. Next, I provide results on the distribution of payoffs among traders and then I perform some comparative statics. Finally, I discuss efficiency under incomplete information. All proofs are in appendix B. 5.1 Final Customers and Intermediaries In a given equilibrium, a trader is active if he takes at least one action with positive probability. Player 1 is always active and makes a strictly positive expected profit. Active traders other than the initial owner can be divided in two classes: final customers and intermediaries. Final customers are active traders who only get offers at price one. This implies that final customers never acquire the object if they have value zero, and their payoff is always zero. Intermediaries are active traders who, at some point in the game, get an offer such that they can buy the object, even if they have value zero, in order to resell it in their neighborhood. In contrast to the complete information case, under asymmetric information an intermediary with value one obtains a positive expected profit. An intermediary with value zero, instead, makes zero profit, because no seller will ever ask him to pay a price below his equilibrium resale value (see observations 3 and 6) The analysis could be extended to equilibria where sellers mix within offers along the equilibrium path without much difficulties. However an additional machinery would be needed to present the results. 22 The fact that only intermediaries extract a positive rent is a consequence of the binary value assumption. In a more general setting, final customers would be also able to extract a lower but still positive information rent, as standard in asymmetric information environments. 14

15 To summarize, the following Table 1 reports the sign of the interim payoff (value one in the second and value zero in the third column) and the ex-ante payoff (fourth column) for player 1, final customers and intermediaries. Table 1: Interim and Ex-ante Equilibrium Payoffs Roles (π i > 0) U i (1) U i (0) U i Initial Owner 1 0 > 0 Intermediary Final Customer The distinction between intermediaries and final customers is exhaustive for each equilibrium of network trading games, as stated in the following proposition. Proposition 3. In every equilibrium of a network trading game G, π, T, every active player is either the initial owner, an intermediary or a final customer. The idea of as follows. Whenever a player is not an intermediary, he will only obtain offers that he will accept exclusively if he has value one. However, if this is the case, the last of these offers must be at price one, otherwise that seller could improve his profit. Therefore, anticipating this, previous sellers will also make offers at price one. Whether a player is either active or not and whether he is either a final customer or an intermediary are determined endogenously in equilibrium and will depend on the complex interaction of the exogenous variables. 23 While it is difficult to derive simple conditions that identify the roles of traders in arbitrary trading network games, there are two classes of traders, isolated traders and bottleneck traders, whose equilibrium role in the trading network can be more easily characterized. A player i 1 is an isolated trader of G if he is connected to only one player, that is ij E for only one j N \ i. For isolated traders I can establish the following proposition. Proposition 4 (Isolated Traders). In every equilibrium of any game G, π, T, an isolated trader i with π i > 0 is either a final customer or an inactive trader. 23 The fact that players with intermediating roles arise endogenously is in contrast with other models of intermediation, where intermediaries are exogenously assigned to their role, as for example in Rubinstein and Wolinsky (1987), and Blume et al. (2009). 15

16 The idea is that players who are at the perimetry of the network will never be useful to intermediate the good to some other area of the network. Therefore no trader will have no interest in selling the object to them at prices lower than one. A bottleneck trader is a trader i 1 whose presence is necessary for the network to remain connected. Formally, let G i be the network obtained by removing node i and all his incident edges from G. 24 Player i is a bottleneck trader in G if the network G i is not connected. Now, let G 1 i denote the largest connected subgraph of G i which includes player 1. Let G i = G G 1 i be the connected subgraph of G which contains i and is obtained by deleting all vertices in G 1 1 and their incident edges. 25 Call Ũi(0) the maximum equilibrium payoff of i with v i = 0 in G i, π, T, assuming that i is the initial owner of an object. 26 Proposition 5 (Bottleneck Traders). Consider a network trading game G, π, T, and suppose that i is a bottleneck trader of G. If Ũi(0) > π i, then in every equilibrium where i is an active player, i is an intermediary. Proposition 5 states that a sufficient condition for an active trader to be an intermediary is that the profit that he can extract from reselling to the part of the network for which he provides monopolistic access is greater than his own expected value. When this is the case, a player selling to a bottleneck trader i will always prefer to demand a price equal to the resale value of i, and have this offer accepted for sure, rather than asking a price of one, and selling only if i has value one. The concept of structural hole, introduced by Burt (1992), refers to the absence of connections within two groups of agents in a social network. Burt s argument is that individuals who fill structural holes, by offering connection between otherwise separated groups, obtain important advantages, in economic and social terms. My analysis provides a foundation for such advantage, explaining how individuals who are essential for connecting a valuable part of the trading network to the initial owner may extract a larger rent than other individuals Write G i = (N \ {i}, E i ), where E i = {{j, j } E : j, j i}. In general, a network G is a subset of G if it is obtained from G by removing a set of players and their incident edges. 25 In example 1, player 2 is a bottleneck trader. G 1 i = ({1, 3}, {13}) and Gi = ({2, 4}, {24}). 26 Observe that T is always defined specifically for the game under consideration. 27 The analysis of structural holes in Goyal and Vega-Redondo (2007) adopts a surplus sharing rule that provides an exogenous advantage to players who have an intermediating role. Hence, they consider a network formation game and focus on whether equilibrium networks include players who fill structural holes. 16

17 While bottleneck and isolated traders do not exhaust all possible types of players in arbitrary trade networks, a player is either an isolated or a bottleneck trader when the trading network is a tree, that is a graph where every pair of players is connected via a unique path. The implications of Proposition 5 are illustrated in the following example. 6 Π 6 Π 3 Π 3 Π 8 Π 8 Π 5 Π 5 Π 2 Π 2 Π 1 Π 1 Π 7 Π 7 Π 4 Π 4 Π Figure 2: Trading Network in Examples 2 and 3 Example 2 (Isolated Traders and Bottleneck Traders in a tree). Assume that players 1, 2,..., 8 are located in the network depicted in Figure 2, that T 7, player 1 is the initial owner, and 1 > π > 0. The set of isolated traders is {3, 4, 6, 7, 8}, while {2, 5} is the set of bottleneck traders. Note that Ũ2(0) = 1 (1 π) 5 > π and Ũ5(0) = 1 (1 π) 3 > π. Therefore if player 2 and 5 are active players in equilibrium, then they are intermediaries. It is not difficult to see that in any network that is a tree, if the number of rounds is sufficiently large, (i) all players connected to player 1 will be active, and (ii) a bottleneck trader will be active if and only if all other bottleneck traders in the unique path going from player 1 to him are active intermediaries In example 2, trader 2 is active because he is connected to trader 1. Therefore trader 5 is also active. 17

18 5.2 Payoff Ranking for Intermediaries This section investigates how the location of an intermediary in the trading chain affect his payoff, when he has value one. 29 By trading chain I mean the ordered sequence of intermediaries that receive offers with positive probability, at prices equal to their resale values. 30 The question is whether an intermediary who makes his first appearance early in the trading chain achieves a higher payoff than an intermediary who appears later, or viceversa. Two countervailing effects are present. First, intermediaries who obtain offers in later rounds will get them with a lower probability than intermediaries who get offers earlier in the game. Second, offers received by intermediaries in later rounds will be at lower prices, as formally stated in the next proposition. Proposition 6 (Decreasing Prices). Consider a trading network game G, π, T. Let i t and i t+k indicate two intermediaries along the trading chain, selling the object in round t and t + k, with k N. Let p t 1 and p t+k 1 be the price that i t and i t+k pay for the object, if they receive offers, in rounds t 1 and t + k 1. In equilibrium, p t 1 p t+k 1 holds with equality if and only if no other player accepts with positive probability an offer from round t to t + k at a price greater than p t+k The price of the object decreases over time because, as rounds of trade take place, it becomes known that there are fewer traders who are potentially interested in consuming the object. This, in turn, reduces the resale value of the object, which is the price that intermediaries are asked to pay. More formally, assume that i t and i t+k are two consecutive intermediaries in the trading chain and let X(i t ) = {x t, x t+1,..., i t+k } be the ordered set of players to which i t makes offers from round t to t + k 1. Let α j indicate the probability that player j X(i t ) accepts his offer. The price at which an intermediary i t acquires the 29 Recall that final customers and intermediaries with value zero always make zero profit. 30 Recall that we consider equilibria where sellers adopt a pure strategy along the equilibrium path and therefore the chain is well defined. 31 Suppose that transaction costs were present and equal to τ for each edge of the network. In this case the relation within prices in round t 1 and t + k 1 would be p t 1 + τ p t+k 1. That is, transaction costs tend, ceteris paribus, to reduce earlier prices as the total amount of the expected transaction costs that remains to be paid must be decreasing in time. 18

19 good in round t 1 is equal to his resale value, that is: p t 1 = α x tp t + (1 α x t)α x t+1p t p t+k 1 j X(i t )\i t+k (1 α j ) Because the seller is rational the sequence of prices must be weakly decreasing and we can conclude that p t 1 p t+k 1. The decrease in price and the reduced probability of obtaining an offer are countervailing forces, but the second dominates the first. In particular, the decrease in price is sufficient to compensate intermediaries for the reduced probability of getting an offer induced by the event that some final customers, who get earlier offers, have value one and consume the object. Nevertheless, it does not compensate the later intermediary for the reduction in probability due to the event that other intermediaries who intervene earlier in the trading chain may themselves consume the object if they have value one. 32 However, it it not possible to conclude from this observation alone that a clear-cut interim payoff ranking exists for intermediaries with value one. In fact, an intermediary could also receive an offer earlier in the game (i.e. before the round in which he acquires the object at his resale value), at a price that he is supposed to accept only if he has value one. This early offer will not change his interim payoff, because a rational seller will keep him indifferent between accepting and rejecting. However, it will reduce the payoff of intermediaries who have earlier positions than him in the trading chain, but receives their offers after he has obtained the early offer. Therefore, when value are known but the game has not started yet, a payoff ranking among intermediaries with value one participating in the trading chain can not be established in general. Instead, when all agents are ex-ante identical, an offer made early to some intermediary will at most equalize his payoff to the level of that of the intermediary preceding him in the trading chain. Therefore, as formally stated in the following proposition, it is possible to establish that intermediaries who are earlier in the trading chain will obtain payoff greater or, at worst, equal to the intermediaries that come later This phenomenon is easy to see if two intermediaries acquire the object one immediately after the other with no other players receiving offers in between. They must be paying the same price, but the second one is worse off than the first, as he only gets the good if the previous one has value zero. 33 The result is robust to the presence of small transaction costs and the introduction of discount rates 19

20 Proposition 7 (Payoff Ranking for Intermediaries). Consider a network trading game G, π, T where π i = π j > 0 for all i, j N. Take any equilibrium of the game and let i t and i t+k indicate two intermediaries along the equilibrium path, selling the object for the first time in round t and t + k, with k N, then U i t(1) U i t+k(1). This result shows that the heterogeneity in outcomes generated by the network structure is not limited to that arising from the different roles of traders in the network (i.e. intermediaries or final customers). The following example illustrates the result. Example 3 (Intermediaries Payoff ranking in a tree). Consider again the environment in example 2. It can be checked that the equilibrium path is the following: Player 1 first asks a price p 1 = 1 (1 π) 5 to intermediary 2. If v 2 = 0, player 2 ask a price of one to final customer 3 and 4, and then asks a price p 4 = 1 (1 π) 3 to intermediary 5. In fact, if player 5 has value zero, he asks a price of one to his final customers 6 and 7 and 8. Therefore, conditional on both having value one, the relation between the interim payoff of 2 and 4 is U 2 (1) = (1 π) 5 > U 4 (1) = (1 π) 6. In general, whenever the network is a tree it is possible to rank the payoff between any two traders who are on the same path to the initial owner, even if traders are ex-ante heterogeneous. In this case, in fact, it is not possible for an intermediary who is later in the trading chain to receive an offer before an intermediary who comes earlier. Therefore, for any two given intermediaries with non zero expected value lying on the same path from the initial owner, we can conclude that the one closer to the initial owner will obtain a strictly higher payoff, when he has value one, than the one who is more far away. 5.3 Comparative Statics It is now possible to perform some comparative statics on how equilibrium payoffs change in response to changes in the exogenous variables. Since my main focus is on the effects of the network structure on trading outcomes, I will examine changes in connectivity only. I say that the network G = (N, E ) is more connected than G = {N, E} if N = N and E E. First, consider the initial owner of the object. close to one. Furthermore, when players are asymmetric, it is always possible to make the proposition valid for traders who are sufficiently far away in the trading chain. 20

21 Proposition 8 (Comparative Statics: Initial Owner). Let G, π, T and G, π, T be two games which differ only because G is more connected than G. Then for every equilibrium of G, π, T there exists an equilibrium of G, π, T where the initial owner achieves a higher or equal expected payoff. 34 The intuition behind the result is that an increase in connectivity accentuates downstream competition between intermediaries. In contrast, there is no analogous result for traders other than the initial owner. In particular, a change in connectivity could have both positive and negative effects on them. The reason is that an increase in connectivity that increases the resale value of a player will be beneficial to him if it changes him from being a final customer to being an intermediary, but it will have a negative effect if he is already an intermediary, because it will increase the price that he pays for the object. This is illustrated by the following example. Consider the two networks depicted in Figure 3 and let the label new on an edge indicate the extra edge that is added to the graph. Consider the equilibrium outcome before and the after the addition of the new edge. In case (a) the equilibrium utility of player 2 when v 2 = 1 decreases from U 2 (1) = 1 π m to U 2 (1) = (1 π m ) 2 when the extra edge is added. That is, player 2 acquires the good at a higher price after the introduction of the new edge. Instead, in case (b) the equilibrium utility of player 2 when v 2 = 1 increases from U 2 (1) = 0 to U 2 (1) = 1 π h when the extra edge is added. In this case player 2 is a final customer initially and becomes an intermediary after the introduction of the new edge, because 1 prefers to route the good via the lowest expected value trader. The effect of an increase in connectivity on total welfare can sometimes be negative. This phenomenon has been referred to in the literature on transportation networks as Braess paradox. To see this point consider the two networks depicted in figure 4. In case (a) the equilibrium is ex-post efficient before the introduction of the new edge, while it is inefficient thereafter. In fact player 1 will only offer the object at price 1 to players 2 and 4 and player 3 will never receive an offer, even if he happens to be the only player in the network with value one. In case (b) the introduction of a new edge makes the outcome of the trading game ex-post efficient, as in the new equilibrium all players receive offers with positive probability. 34 This result will not hold if there are large transaction costs or the number of rounds is not sufficiently large. 21